mdlp/CLAUDE.md

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project Overview

This is a C++ implementation of the MDLP (Minimum Description Length Principle) discretization algorithm based on Fayyad & Irani's paper. The library provides discretization methods for continuous-valued attributes in classification learning.

Build System

The project uses CMake with a Makefile wrapper for common tasks:

Common Commands

• make build - Build release version with sample program
• make test - Run full test suite with coverage report
• make install - Install the library

Build Configurations

• Release: Built in build_release/ directory
• Debug: Built in build_debug/ directory (for testing)

Dependencies

• PyTorch (libtorch) - Required dependency
• GoogleTest - Fetched automatically for testing
• Coverage tools: lcov, genhtml

Code Architecture

Core Components

1. Discretizer (src/Discretizer.h/cpp) - Abstract base class for all discretizers
2. CPPFImdlp (src/CPPFImdlp.h/cpp) - Main MDLP algorithm implementation
3. BinDisc (src/BinDisc.h/cpp) - K-bins discretization (quantile/uniform strategies)
4. Metrics (src/Metrics.h/cpp) - Entropy and information gain calculations

Key Data Types

• samples_t - Input data samples
• labels_t - Classification labels
• indices_t - Index arrays for sorting/processing
• precision_t - Floating-point precision type

Algorithm Flow

1. Data is sorted using labels as tie-breakers for identical values
2. MDLP recursively finds optimal cut points using entropy-based criteria
3. Cut points are validated to ensure meaningful splits
4. Transform method maps continuous values to discrete bins

Testing

Tests are built with GoogleTest and include:

• Metrics_unittest - Entropy/information gain tests
• FImdlp_unittest - Core MDLP algorithm tests
• BinDisc_unittest - K-bins discretization tests
• Discretizer_unittest - Base class functionality tests

Running Tests

make test  # Runs all tests and generates coverage report
cd build_debug/tests && ctest  # Run tests directly

Coverage reports are generated at build_debug/tests/coverage/index.html.

Sample Usage

The sample program demonstrates basic usage:

build_release/sample/sample -f iris -m 2

Development Notes

• The library uses PyTorch tensors for efficient numerical operations
• Code follows C++17 standards
• Coverage is maintained at 100%
• The implementation handles edge cases like duplicate values and small intervals
• Conan package manager support is available via conanfile.py